Over the past thirty years, researchers have made real progress engineering artificial intelligence (AI) into commercial systems. Automatic translation, speech understanding, reasoning with constraints, logic, game playing, image recognition, and industrial robotics are all well on their way to being mastered. But one benchmark problem still exposes some of AI’s deepest remaining challenges: reproducing the simple act of reading a textbook, understanding the material inside, and answering questions about it.
Why is this so difficult for computers? After all, learning new things, working though their implications, and answering questions are all second nature to us—we do them so easily that we rarely stop to consider the mechanisms involved. And computers certainly have enough raw power to do the job; modern search engines can sift through the Web in less than a second and deliver pages that match our search terms, ranked in order of usefulness. Nevertheless, getting a computer to answer ordinary questions of the sort commonly found on high school exams and answered by millions of students is extremely challenging to replicate.
The problem has to do with the nature of human knowledge itself. Knowledge is often thought of as a large collection of facts, like multiplication tables or lists of chemical properties. Indeed, existing artificial intelligence technologies can answer questions that depend only on simple facts. (“How many chromosomes does a blue jay have?”). But the most important elements of human knowledge involve much more sophisticated constructions. Even cut-and-dried knowledge includes rough statements of causality (“Too little sunlight can lead to stunted plants”), generality (“Most birds can fly”), metaphor (“DNA is like a blueprint”), counter-factuals (“If Earth’s gravity were halved, trees could be twice as tall”), rule knowledge (“If a cell dies, its cell membrane disintegrates”), and prediction (“Mutations should increase in the presence of radioactivity”).
The goal of the Digital Aristotle project is to find ways for computers to grapple with all types of human knowledge, and to manage and manipulate their full range and richness. In order to succeed, it will need to acquire knowledge intelligently, reason through it effectively, and find appropriate answers on a truly massive scale.
Our Project Halo research program is designed to build the systems that can ultimately lead to a functional Digital Aristotle. We began Project Halo several years ago by targeting biology at the level of a high school Advanced Placement course. This subject area served our purpose because it has significant (but not overwhelming) scale, features a set curriculum with accepted tests for competence, and exhibits many of the more challenging types of knowledge. Thus far, we have analyzed standard biology textbooks line by line in order to categorize each type of knowledge they contain. Now we are working on ways to encode these types of knowledge into Project Halo’s computers, merge them with the knowledge that is already there, and keep everything in a form that will allow our various reasoning systems to respond with the correct answer to a user’s questions.
The basic challenge in all this work is its pervasive brittleness. Many tough problems in computer knowledge encoding and reasoning have been successfully addressed at a small scale in a laboratory. But when these efforts scale up—even to the amount of knowledge in a single biology textbook—they break. Furthermore, the individual approaches are often incompatible with one another, and so current AI systems can’t match people’s fluid shifts between different ways of using their knowledge.
The international Project Halo team has made considerable progress in our research. We believe that by 2015 we’ll be able to build a system that includes most of the knowledge required to answer Advanced Placement–level biology questions. This system, in the form of a tabletlike Halobook, will constitute an important step in our pursuit of the Digital Aristotle. Nevertheless, difficult challenges remain; the ultimate solution will require many more breakthroughs. Here are ten areas of knowledge representation that are currently formidable for machines to handle and are of interest to Project Halo, grouped into three tiers of difficulty:
I. DIFFICULTY TIER 1: PROMISING APPROACHES STILL HALFWAY AT MOST TO A ROBUST SOLUTION
· Human language is powerful and complex. There are many ways of saying the same thing, and many different things communicated in every sentence. For a machine to process the full range of human language, it must “understand” and react appropriately to a huge variety of potential expression. Many promising techniques are being developed using both manual and automated analysis of language, including statistical studies of massive data sets drawn from the Web. The intersection of language and knowledge is an area that we have great interest in and are actively pursuing in Project Halo, dealing with the full range of linguistic expression.
· Visual/spatial learning and reasoning. Can seventeen suitcases fit inside the trunk of a typical car? What about an open umbrella? Can a jetliner land on a sidewalk? What information is represented in a diagram? How does DNA uncoil? Humans perform rough-and-ready spatial and visual reasoning tasks and visual simulations with ease. While the computational geometry that is needed for navigation, manufacturing, and architecture exists and is commercially available, progress has been much slower in dealing with the kind of intuitive geometry that we routinely use every day. Project Halo does not currently focus in this area but welcomes new ideas.
· Knowledge about actions, causality, and simulation. If a cup is on a table in a room, and a person enters the room, the cup is unaffected; it will still be on the table. But other things do change as a result of this action: the person will no longer be outside the room; the person’s body and clothes will be in the room; the room will no longer be empty, and so on. Humans effortlessly perform mental simulations in their heads, both in a “forward” direction to predict how events might play out, and in a “backward” direction to identify likely causes. Computationally, however, this is a difficult and long-standing problem for AI. Reasoning about actions, change, and causality is extremely complex, especially when an action’s effects are uncertain and have indirect consequences. The best current solutions are found in business processes, automatic planning, and robotics, but they tend to be highly customized and difficult to apply to new areas. Project Halo has made substantial progress in general reasoning about processes and actions.
· Handling pervasive uncertainty and vagueness. Much of our knowledge is uncertain, vague, and approximate, yet we have a remarkable ability to draw conclusions and act. After listening to the weather forecast, a person who knows that it might rain can make contingency plans. People can read vague statements (“John is fairly tall”), approximations (“The human genome contains about 23,000 genes”), or statements with exceptions (“All birds can fly”) and still draw useful conclusions despite the imprecision. Classic techniques of statistics can already address these issues in selected domains, and may yet work in general, but progress has been slow. Project Halo has made some advances in this area, which remains an important focus of our research.
II. DIFFICULTY TIER 2: RESEARCH THAT IS STILL PRELIMINARY AND EXPERIMENTAL
· Unstated and implicit knowledge in language. Human language is full of ambiguity and gaps in knowledge that a reader or listener must interpret correctly. Take, for example, the statement “A teaspoon of salt is dissolved in water.” Is it the teaspoon or the salt that is dissolved? Is the teaspoon made of salt? Humans use knowledge to instantly resolve such ambiguities, while machines struggle. If we read that “acids can cause some dyes to change color,” we immediately assume that the acid and dye must be in contact, although it’s not explicitly stated. To accurately understand statements like these, our brains make use of a rich interplay between textual and background knowledge. For a computer to have full language understanding, it needs to overcome this critical problem.
· Evolving knowledge. Acquiring new knowledge is not simply a matter of memorization. New knowledge always needs to be “fitted in” with existing knowledge in a way that is coherent. For example, if you learn a simple model of how cells divide, and then come across a more complex description, you recognize that you need to align the two, which modifies your original understanding. Perhaps what you originally thought of as a single process now needs to be revised and conceptualized as two linked ones. This process of maintaining, revising, and expanding existing knowledge is critical for large-scale systems like the Digital Aristotle. Simple, specialized techniques for doing this exist, but a fully automated solution seems decades away.
· Contradictions, fragility, and handling messy knowledge. While knowledge bases for small-and medium-scale artificial intelligence systems can be fully debugged, knowledge bases above a certain size inevitably become “messy” with errors, inconsistencies, gaps, and contradictions. As the volume of available data and knowledge grows, AI systems need to both effectively debug artifacts and to continue to reason in a robust, sensible way. This challenge becomes particularly significant in Web-scale systems, where sources of knowledge and data may be geographically, culturally, and temporally diverse. A variety of new techniques exist here, from fancy new logics to systems inspired by Web search technology, but they are still experimental. Project Halo is actively working on reasoning techniques that will handle several specific kinds of conflict and contradiction, but general solutions have been elusive.
· Commonsense reasoning. A vast amount of our understanding draws on general, commonsense knowledge and rules of thumb. For example, if you are told that “carbon dioxide is a raw material for photosynthesis,” you readily infer that carbon dioxide is used in photosynthesis, that it is required and also consumed. You can draw these inferences because you understand these general notions (“raw material,” “require,” “consume”) and the relationships between them. Commonsense knowledge provides great flexibility in human question-answering and reasoning, but correctly applying it in machines is a major challenge. A range of systems now exist, from those that attempt to use the Web to systems like Cyc (www.cyc.com), which are mostly human-authored. But while computers can demonstrate examples of commonsense reasoning, their ability to reliably acquire and use this type of knowledge at the scale required for a Digital Aristotle remains unproven. Project Halo is working to find solutions.
III. DIFFICULTY TIER 3: SOME OF THE TOUGHEST REMAINING CHALLENGES IN AI
· Applying knowledge in new contexts. Humans apply their knowledge in new contexts, constructing innovative and often novel ideas. For example, when a high school student designs an experiment to validate a chemical principle, she is capable of managing her existing knowledge about actions and objects to assemble it into a suitable sequence. We do the same thing when we imagine fictional situations, using what we know in new ways and applying it to new contexts. This ability to manipulate existing knowledge in complex and original ways remains a major challenge for computers. Very little exists in this area beyond preliminary research.
· Metaphor and analogy. When confronted with something new, people frequently draw on and adapt what they already know. For example, one biology text states, “Microtubules in the cell are like miniature springs.” The analogy prompts a reader to draw on existing knowledge to understand how microtubules expand and contract, yet avoid the conclusion that microtubules are likely made of metal. This skill requires identifying, mapping, and selectively adapting existing mental models to new tasks for which the model was never intended. Such a process remains almost impossible to automate. Very little exists in this area beyond preliminary research.
For each of these types of knowledge representation (and several more), Project Halo is actively seeking solutions worldwide. If you have serious technical ideas in these areas, please contact us at ideas@projecthalo.com.
*Little did I know that the “machine” on the cover was in fact a hollow mock-up, subbed in at the last minute after the genuine Altair prototype was delayed in shipping by a Railway Express strike.
*Between the paper tape era and the popularization of floppy disks, audiocassettes had a brief run in the midseventies as a leading storage device for microcomputers.
*Scientific notation is a simplified way to handle very small or very large numbers using coefficients and exponents. For example: 83,700,000 = 8.37 × 107; 0.0072 = 7.2 × 10−3.
*The opening credits embedded in our BASIC were as follows: “Paul Allen wrote the non-runtime stuff. Bill Gates wrote the runtime stuff. Monte Davidoff wrote the math package.”
*The theft of Altair BASIC foreshadowed the wholesale piracy of copyrighted material that plagues the entertainment industry today. Once a song or movie or piece of software was reduced to binary bits, it became easy to copy, even more so with the ascendance of the Internet.
*Earlier that summer, when we were still licensing 86-DOS under a nonexclusive contract, Brock was approached by Eddie Currie on behalf of Lifeboat Associates. As Currie tells the story, he offered Brock $250,000 for any rights to 86-DOS that Microsoft didn’t control. Brock chose instead to stay with us. He didn’t want to antagonize Bill or lose his long-term, cut-rate access to our software.
*As minicomputers were undercut by ever more powerful microprocessors, and by the PC in particular, DEC went into a fatal tailspin and was acquired by Compaq in 1998.
*The public contribution would come from the interests who stood to benefit most, via an extended county hotel tax and increased parking and admission taxes at the stadium, along with new lottery games and a state sales tax credit that reflected the team’s economic value to Seattle.